Segmentation and feature analysis of medical images for cell abnormalities and cancer detection

Current advances in image capture devices have resulted in significant steps forward in medical applications. Details of human anatomies can be clearly observed, even at the cellular level. Such advances in imaging techniques not only allow clinicians to study human biometrics clearly, the visual information can be easily recorded for future reference and study. Digital image capture can also provide the clinician with a non-invasive alternative to performing medical diagnosis for diseases such as cancer. However, a manual visual analysis of the large quantity of medical image data tends to be laborious and time consuming and can be highly subjective due to inter-observer variability. This has motivated researchers to utilise computational power to develop a computer-aided diagnosis (CAD) system to assist clinicians in medical diagnosis. One of the major hurdles researchers face when designing an automatic CAD system for medical image analysis is that the segmentation algorithms are not versatile enough to correctly delineate the region of interest from the images captured in different settings. For instance, ambient light may result in different skin tone variations in dermatography for detecting skin cancer. Furthermore, classification algorithms require rigorous development to improve the accuracy of the diagnosis. In this thesis, we study cell splitting behaviour in the cellular life cycle. We introduce cell deformation features to characterise changes in the shape and size of the cell during the mitosis process, and key stages of the cell cycle can be determined. The outcomes of this research can be used to monitor any abnormal behaviour in the cell cycle. We also study blood smear samples and extract visual features to identify abnormal red blood cells. We introduce a hierarchical structure network to model the human visual cortex to process shape and texture information in a cascade fashion. The proposed classifier can successfully determine four major types of abnormal red blood cells. Skin dermatography images are also studied. We introduce the 3D colour constancy algorithm to overcome the problem of skin tone variations. Type-2 fuzzy algorithms are applied to determine the optimum threshold level to delineate the cancerous region. We apply the hierarchical network structure developed earlier to develop an algorithm to predict the existence of skin melanoma. The proposed algorithm achieves a precision rate of 93.51%, which outperforms the benchmark MLP predictor. x We continue our cancer research by studying radiation therapy (RT) for cancer treatment. We use the colour information in the radiation dosage distribution in the RT plans. The proposed algorithm quantizes and normalises the dosage maps and determines the relative dosage levels at the target site and surrounding vital organs. We propose a quality assessment evaluation scheme for RT which includes the medical condition of the patient. The proposed QA scheme determines the best treatment plans to minimise the radiation damage to organs which may have been weakened by other illnesses.

Submission note: A thesis submitted in total fulfilment of the requirements for the degree of Doctor of Philosophy to the Department of Computer Science and Computer Engineering, School of Engineering and Mathematical Sciences, Faculty of Science, Technology and Engineering, La Trobe University, Melbourne.